Longitudinal Electronic Health Research Articles

Discovering patient groups in sequential electronic healthcare data using unsupervised representation learning

IntroductionUnsupervised feature learning methods inspired by natural language processing (NLP) models are capable of constructing patient-specific features from longitudinal Electronic Health Records (EHR).DesignWe applied document embedding algorithms to real-world paediatric intensive care (PICU) EHR data to extract patient-specific features from 1853 patients’ PICU journeys using 647 unique lab tests and medication events. We evaluated the clinical utility of the patient features via a K-means clustering analysis.ResultsWe trained a document embedding model under a unique evaluation pipeline and obtained latent patient feature vectors for all 1853 patients. We performed unsupervised clustering to the patient vectors as a downstream analysis and obtained 5 distinct clusters via hyperparameter optimisation. Significant variations (p<0.0001) within both patient characteristics and surgery intervention and diagnostic profiles were detected.ConclusionThe K-means clustering results demonstrated the clinical utilities of the patient-specific features learned from the embedding algorithms. The latent patient features obtained via the embedding process enabled direct applications of other machine learning algorithms. Future work will focus on utilising the temporal information within EHR and extending EHR embedding algorithms to develop personalised patient journey predictions.

Development of Attention-based Prediction Models for All-cause Mortality, Home Care Need, and Nursing Home Admission in Ageing Adults in Spain Using Longitudinal Electronic Health Record Data

Predicting health-related outcomes can help with proactive healthcare planning and resource management. This is especially important on the older population, an age group growing in the coming decades. Considering longitudinal rather than cross-sectional information from primary care electronic health records (EHRs) can contribute to more informed predictions. In this work, we developed prediction models using longitudinal EHRs to inform resource allocation. In this study, we developed deep-learning-based prognostic models to predict 1-year and 5-year all-cause mortality, nursing home admission, and home care need in people over 65 years old using all the longitudinal information from EHRs. The models included attention mechanisms to increase their transparency. EHRs were drawn from SIDIAP (primary care, Catalonia (Spain)) from 2010-2019. Performance on the test set was compared to that from baseline models using cross-sectional one-year history only. Data from 1,456,052 individuals over 65 years old were considered. Cohen’s kappa obtained using longitudinal data was 3.4-fold (1-year all-cause mortality), 10.3-fold (5-year all-cause mortality), 1.1-fold (5-year nursing home admission), and 1.2-fold (5-year home care need) higher than that obtained by the one-year history baseline models. Our models performed better than those not considering longitudinal data, especially when predicting further into the future. However, nursing home admission and home care need in the long term were harder to predict, suggesting their dependence on more abrupt changes. The attention maps helped to understand the predictions, enhancing model transparency. These prediction models can contribute to improve resource allocation in the general population of aging adults.

A Self-Supervised Graph Neural Network to Identify Temporal Phenotypes of End-Stage Renal Disease Using Longitudinal Electronic Health Records

A Self-Supervised Graph Neural Network to Identify Temporal Phenotypes of End-Stage Renal Disease Using Longitudinal Electronic Health Records

Multimodal deep learning for predicting in-hospital mortality in heart failure patients using longitudinal chest X-rays and electronic health records.

Amid an aging global population, heart failure has become a leading cause of hospitalization among older people. Its high prevalence and mortality rates underscore the importance of accurate mortality prediction for swift disease progression assessment and better patient outcomes. The evolution of artificial intelligence (AI) presents new avenues for predicting heart failure mortality. Yet current research has predominantly leveraged structured data and unstructured clinical notes from electronic health records (EHR), underutilizing the prognostic value of chest X-rays (CXRs).This study aims to harness deep learning methodologies to explore the feasibility of enhancing the precision of predicting in-hospital all-cause mortality in heart failure patients using CXRs data.We propose a novel multimodal deep learning network based on the spatially and temporally decoupled Transformer (MN-STDT) for in-hospital mortality prediction in heart failure by integrating longitudinal CXRs and structured EHR data. The MN-STDT captures spatial and temporal information from CXRs through a Hybrid Spatial Encoder and a Distance-Aware Temporal Encoder, ultimately fusing features from both modalities for predictive modeling. Initial pre-training of the spatial encoder was conducted on CheXpert, followed by full model training and evaluation on the MIMIC-IV and MIMIC-CXR datasets for mortality prediction tasks. In a comprehensive view, the MN-STDT demonstrated the best performance, with an AUC-ROC of 0.8620, surpassing all baseline models. Comparative analysis revealed that the AUC-ROC of the multimodal model (0.8620) was significantly higher than that of models using only structured data (0.8166) or chest X-ray data alone (0.7479).This study demonstrates the value of CXRs in the prognosis of heart failure, showing that the combination of longitudinal CXRs with structured EHR data can significantly improve the accuracy of mortality prediction in heart failure. Feature importance analysis based on SHAP provides interpretable decision support, paving the way for potential clinical applications.

Early detection of heart failure using in-patient longitudinal electronic health records.

Heart Failure (HF) is common, with worldwide prevalence of 1%-3% and a lifetime risk of 20% for individuals 40 years or older. Despite its considerable health economic burden, techniques for early detection of HF in the general population are sparse. In this work we tested the hypothesis that a simple Transformer neural network, trained on comprehensive collection of secondary care data across the general population, can be used to prospectively (three-year predictive window) identify patients at an increased risk of first hospitalisation due to HF (HHF). The model was trained using routinely-collected, secondary care health data, including patient demographics, A&E attendances, hospitalisations, outpatient data, medications, blood tests, and vital sign measurements obtained across five years of longitudinal electronic health records (EHRs). The training cohort consisted of n = 183,894 individuals (n = 161,658 age/sex-matched controls and n = 22,236 of first hospitalisation due to HF after a three-year predictive window). Model performance was validated in an independent testing set of n = 8,977 patients (n = 945 HHF patients). Testing set probabilities were well-calibrated and achieved good discriminatory power with Area Under Receiver Operating Characteristic Curve (AUROC]) of 0.86, sensitivity of 36.4% (95% CI: 33.33%-39.56%), specificity of 98.26% (95% CI: 97.95%-98.53%), and PPV of 69.88% (95% CI: 65.86%-73.62%). At Probability of HHF ≥ 90% the model achieved 100% PPV (95% CI: 96.73%-100%) and sensitivity of 11.7% (95% CI: 9.72%-13.91%). Performance was not affected by patient sex or socioeconomic deprivation deciles. Performance was significantly better in Asian, Black, and Mixed ethnicities (AUROC 0.932-0.945) and in the 79-86 age group (AUROC 0.889). We present the first evidence that routinely collected secondary care health record data can be used in the general population to stratify patients at risk of first HHF.

Study protocol: Comparison of different risk prediction modelling approaches for COVID-19 related death using the OpenSAFELY platform

On March 11th 2020, the World Health Organization characterised COVID-19 as a pandemic. Responses to containing the spread of the virus have relied heavily on policies involving restricting contact between people. Evolving policies regarding shielding and individual choices about restricting social contact will rely heavily on perceived risk of poor outcomes from COVID-19. In order to make informed decisions, both individual and collective, good predictive models are required. For outcomes related to an infectious disease, the performance of any risk prediction model will depend heavily on the underlying prevalence of infection in the population of interest. Incorporating measures of how this changes over time may result in important improvements in prediction model performance. This protocol reports details of a planned study to explore the extent to which incorporating time-varying measures of infection burden over time improves the quality of risk prediction models for COVID-19 death in a large population of adult patients in England. To achieve this aim, we will compare the performance of different modelling approaches to risk prediction, including static cohort approaches typically used in chronic disease settings and landmarking approaches incorporating time-varying measures of infection prevalence and policy change, using COVID-19 related deaths data linked to longitudinal primary care electronic health records data within the OpenSAFELY secure analytics platform.

PheWAS analysis on large-scale biobank data with PheTK.

With the rapid growth of genetic data linked to electronic health record data in huge cohorts, large-scale phenome-wide association study (PheWAS) have become powerful discovery tools in biomedical research. PheWAS is an analysis method to study phenotype associations utilizing longitudinal electronic health record (EHR) data. Previous PheWAS packages were developed mostly with smaller data sets and with earlier PheWAS approaches. PheTK was designed to simplify analysis and efficiently handle biobank-scale data. PheTK uses multithreading and supports a full PheWAS workflow including extraction of data from OMOP databases and Hail matrix tables as well as PheWAS analysis for both phecode version 1.2 and phecodeX. Benchmarking results showed PheTK took 64% less time than the R PheWAS package to complete the same workflow. PheTK can be run locally or on cloud platforms such as the All of Us Researcher Workbench (All of Us) or the UK Biobank (UKB) Research Analysis Platform (RAP). The PheTK package is freely available on the Python Package Index, on GitHub under GNU General Public License (GPL-3) at https://github.com/nhgritctran/PheTK, and on Zenodo, DOI 10.5281/zenodo.14217954, at https://doi.org/10.5281/zenodo.14217954. PheTK is implemented in Python and platform independent. Supplementary data are available at Bioinformatics online.

DySurv: dynamic deep learning model for survival analysis with conditional variational inference.

Machine learning applications for longitudinal electronic health records often forecast the risk of events at fixed time points, whereas survival analysis achieves dynamic risk prediction by estimating time-to-event distributions. Here, we propose a novel conditional variational autoencoder-based method, DySurv, which uses a combination of static and longitudinal measurements from electronic health records to estimate the individual risk of death dynamically. DySurv directly estimates the cumulative risk incidence function without making any parametric assumptions on the underlying stochastic process of the time-to-event. We evaluate DySurv on 6 time-to-event benchmark datasets in healthcare, as well as 2 real-world intensive care unit (ICU) electronic health records (EHR) datasets extracted from the eICU Collaborative Research (eICU) and the Medical Information Mart for Intensive Care database (MIMIC-IV). DySurv outperforms other existing statistical and deep learning approaches to time-to-event analysis across concordance and other metrics. It achieves time-dependent concordance of over 60% in the eICU case. It is also over 12% more accurate and 22% more sensitive than in-use ICU scores like Acute Physiology and Chronic Health Evaluation (APACHE) and Sequential Organ Failure Assessment (SOFA) scores. The predictive capacity of DySurv is consistent and the survival estimates remain disentangled across different datasets. Our interdisciplinary framework successfully incorporates deep learning, survival analysis, and intensive care to create a novel method for time-to-event prediction from longitudinal health records. We test our method on several held-out test sets from a variety of healthcare datasets and compare it to existing in-use clinical risk scoring benchmarks. While our method leverages non-parametric extensions to deep learning-guided estimations of the survival distribution, further deep learning paradigms could be explored.

Real-World Treatment Patterns and Clinical Outcomes in Patients With Locally Advanced or Metastatic Urothelial Carcinoma by Eligibility for Maintenance Avelumab

IntroductionFirst-line (1L) platinum-based chemotherapy (PBC) has historically been recommended for patients with locally advanced or metastatic urothelial carcinoma (Ia/mUC). Maintenance avelumab is recommended for patients without disease progression following 1L PBC. Real-world data on the proportion of patients eligible for maintenance avelumab are limited, and outcomes among patients ineligible for maintenance avelumab are uncertain. This study assessed the proportion of patients with la/mUC initiating 1L PBC who were maintenance-avelumab eligible and real-world outcomes following 1L PBC by maintenance-avelumab eligibility. MethodsA retrospective, observational study was conducted using a longitudinal electronic health record–derived database comprising de-identified patient-level structured and unstructured data including adults with Ia/mUC who received ≥1 dose of 1L PBC (April 2020-January 2022). The proportion of patients eligible for maintenance avelumab (real-world stable disease, partial response, or complete response following 1L PBC) was estimated and median overall survival (mOS) assessed for maintenance avelumab–eligible and –ineligible patients. ResultsOf 336 patients with Ia/mUC treated with 1L PBC (55.4% received cisplatin-based treatment 44.6% carboplatin-based treatment); 181 (54%) were maintenance-avelumab eligible; and 138 (41%) maintenance-avelumab ineligible (17 [5%] were nonevaluable). Of 181 maintenance-avelumab–eligible patients, 67 (37.0%; 19.9% of all 1L PBC–treated patients) received maintenance avelumab. mOS (95% CI) among all 1L PBC–treated patients was 15.0 (12.2-19.6) months and among maintenance-avelumab–ineligible patients was 8.0 (6.7-10.3) months; whereas among maintenance-avelumab–eligible patients (including 37% who received maintenance avelumab), mOS was 27.6 (23.4-not reached) months. ConclusionsIn this study, approximately half of 1L PBC–treated patients were maintenance-avelumab eligible, and one-fifth received it. Real-world OS remains short for the overall 1L PBC–treated population. These results support the use of treatment-guideline preferred 1L treatment options that demonstrate survival benefit for all patients with la/mUC, and are available to patients irrespective of their eligibility for cisplatin, or response to PBC.

Abstract 4125887: Aladyn Individual: Dynamic Individual Comorbidity Modeling for Genomic Discovery and Clinical Prediction

Intro: Considering comorbidity patterns for both discovery and prediction of individual disease trajectories is essential for personalized medicine. Existing methods often lack genetic integration and interpretability. We develop an approach suited for dynamic environments that combines genetics with individual and population level-trajectories for enhanced discovery and prediction. Method: Our method employs a dynamic Bayesian hierarchical latent allocation framework to learn and predict disease trajectories. Leveraging longitudinal electronic health care data from 485,698 UK Biobank participants and 287,012 All of Us Research Program participants, we aim to learn individual and population level disease trajectories for 352 distinct conditions. Genetic risks inform the model in two fundamental ways: First, we introduce a genetic warping parameter that recognizes the tendency of an individual to experience events earlier or later within these profiles while retaining the relative order of progression (1a). Second, we introduce a gaussian process to combine individual and population level trends in comorbidity profile prevalence over time (1b), whose mean is augmented by genetic predilection. Finally, we allow for profile weights to update with each new diagnoses, merging prediction and classification. Results: Our results indicate stark disparities in disease progression; individuals at high genetic risk for myocardial infarction experience events nearly a decade earlier than those at low risk (median age of onset 53.7 [53.1-53.9] vs 62.6 years [61.4-63.1]) (1b). Our model dynamically updates and predicts correct diagnoses 79.5% more frequently than existing topic-modeling approaches (1c,d). The dynamic approach adjusts disease profiles by at least 11.7% annually for 50% of individuals, with 65% experiencing over a 10% change in comorbidity profiles. Simulation results confirm the model's superior accuracy [51.4-41%], precision [50.9 vs 39.9%], and recall [48.4 vs 39.2%] compared to fixed-weight approaches. Summary: We offer a novel approach for integrating genetic risk with dynamic modeling of complex disease profiles for joint biologic discovery and personalized clinical prediction over the lifetime.

Precision phenotyping for curating research cohorts of patients with unexplained post-acute sequelae of COVID-19.

Scalable identification of patients with post-acute sequelae of COVID-19 (PASC) is challenging due to a lack of reproducible precision phenotyping algorithms, which has led to suboptimal accuracy, demographic biases, and underestimation of the PASC. In a retrospective case-control study, we developed a precision phenotyping algorithm for identifying cohorts of patients with PASC. We used longitudinal electronic health records data from over 295,000 patients from 14 hospitals and 20 community health centers in Massachusetts. The algorithm employs an attention mechanism to simultaneously exclude sequelae that prior conditions can explain and include infection-associated chronic conditions. We performed independent chart reviews to tune and validate the algorithm. The PASC phenotyping algorithm improves precision and prevalence estimation and reduces bias in identifying PASC cohorts compared to the ICD-10-CM code U09.9. The algorithm identified a cohort of over 24,000 patients with 79.9% precision. Our estimated prevalence of PASC was 22.8%, which is close to the national estimates for the region. We also provide in-depth analyses, encompassing identified lingering effects by organ, comorbidity profiles, and temporal differences in the risk of PASC. PASC precision phenotyping boasts superior precision and prevalence estimation while exhibiting less bias in identifying patients with PASC. The cohort derived from this algorithm will serve as a springboard for delving into the genetic, metabolomic, and clinical intricacies of PASC, surmounting the constraints of prior PASC cohort studies. This research was funded by the US National Institute of Allergy and Infectious Diseases (NIAID).

U.S. weight trends: a longitudinal analysis of an NIH-partnered dataset.

Obesity is a major public health challenge in the U.S. Existing datasets utilized for calculating obesity prevalence, such as the National Health and Nutrition Examination Survey (NHANES) and Behavioral Risk Factor Surveillance System (BRFSS), have limitations. Our objective was to analyze weight trends in the U.S. using a nationally representative dataset that incorporates longitudinal electronic health record data. Using the National Institutes of Health All of Us Research Program (AoU) dataset, we identified patients aged 18-70 years old who had at least two height and weight measurements within a 5-year period from 2008 to 2021. Baseline and most recent BMI values were used to calculate total body weight (%TBW) changes. %TBW change predictors were determined using multivariable linear regression. We included 30,862 patients (mean age 48.9 [ ± 12.6] years; 60.5% female). At the 5-year follow-up, the prevalences of obesity and severe obesity were 37.4% and 20.7%, respectively. The frequency of patients with normal weight or overweight BMI who gained ≥5% TBW at follow-up was 37.8% and 33.1%, respectively. Nearly 24% of the cohort lost ≥ 5% TBW, and 6.5% with severe obesity lost weight to achieve a BMI < 30 kg/m2. In adjusted analyses, male sex (-1.10%, 95% CI [-1.36, -0.85]), non-Hispanic Asian race/ethnicity (-1.69% [-2.44, -0.94]), and type 2 diabetes (-1.58% [-1.95, -1.22]) were associated with weight loss, while obstructive sleep apnea (1.80% [1.40, 2.19]) was associated with weight gain. This evaluation of an NIH-partnered dataset suggests that patients are continuing to gain weight in the U.S. AoU represents a unique tool for obesity prediction, prevention, and treatment given its longitudinal nature and unique behavioral and genetic data.

Unsupervised clustering of longitudinal clinical measurements in electronic health records.

Longitudinal electronic health records (EHR) can be utilized to identify patterns of disease development and progression in real-world settings. Unsupervised temporal matching algorithms are being repurposed to EHR from signal processing- and protein-sequence alignment tasks where they have shown immense promise for gaining insight into disease. The robustness of these algorithms for classifying EHR clinical data remains to be determined. Timeseries compiled from clinical measurements, such as blood pressure, have far more irregularity in sampling and missingness than the data for which these algorithms were developed, necessitating a systematic evaluation of these methods. We applied 30 state-of-the-art unsupervised machine learning algorithms to 6,912 systematically generated simulated clinical datasets across five parameters. These algorithms included eight temporal matching algorithms with fourteen partitional and eight fuzzy clustering methods. Nemenyi tests were used to determine differences in accuracy using the Adjusted Rand Index (ARI). Dynamic time warping and its lower-bound variants had the highest accuracies across all cohorts (median ARI>0.70). All 30 methods were better at discriminating classes with differences in magnitude compared to differences in trajectory shapes. Missingness impacted accuracies only when classes were different by trajectory shape. The method with the highest ARI was then used to cluster a large pediatric metabolic syndrome (MetS) cohort (N = 43,426). We identified three unique childhood BMI patterns with high average cluster consensus (>70%). The algorithm identified a cluster with consistently high BMI which had the greatest risk of MetS, consistent with prior literature (OR = 4.87, 95% CI: 3.93-6.12). While these algorithms have been shown to have similar accuracies for regular timeseries, their accuracies in clinical applications vary substantially in discriminating differences in shape and especially with moderate to high missingness (>10%). This systematic assessment also shows that the most robust algorithms tested here can derive meaningful insights from longitudinal clinical data.

Identification of Novel Genetic Risk Variants Associated with Hidradenitis Suppurativa in an Exome Sequencing Cohort of 92,455 Individuals

Introduction: Hidradenitis suppurativa (HS) is a prevalent and persistent inflammatory skin disorder, lacking a known cure or effective biomarkers for early diagnosis at present. The genetic determinants of HS have not been fully documented, but it is believed to result from a combination of genetic and environmental factors. Methods: To identify relevant HS gene variants in sporadic HS patients, this study utilized longitudinal electronic health records (EHRs) and whole-exome sequencing. DNA exome sequencing data from 92,455 participant samples in the MyCode biobank, linked to Geisinger’s EHR, were analyzed. This cohort included 1,092 HS cases and 91,363 healthy controls. The MyCode EHR has a median longitudinal follow-up of 15 years per participant, with an average of 87 clinical encounters, 687 laboratory tests, and 7 procedures. Results: There were 1,092 (901 females and 191 males) participants aged 14–89 years (median 47 years) with HS (L73.2), indicating a 1.18% prevalence and accounting for a 4.7:1 female-to-male ratio among the individuals presenting for clinical care. γ-secretase complex, syndromic, and autoinflammatory gene variants were assessed. Potential pathogenic variants were identified among 66 individuals in the HS genes studied. Molecularly, the estimated HS variant prevalence was 1:1,400 in the cohort, 12.3% of variant carriers had HS diagnosis in EHR. Conclusions: Using longitudinal EHR data, genomic screening identified HS-associated gene variants in a defined group of sporadic HS patients to augment the clinical diagnosis, particularly in cases of ambiguity. Based on this study, the field of skin disorders can benefit from a personalized approach to HS diagnosis using large-scale sequencing.

High Burden of Ileus and Pneumonia in Clozapine-Treated Individuals With Schizophrenia: A Finnish 25-Year Follow-Up Register Study.

The authors used longitudinal biobank data with up to 25 years of follow-up on over 2,600 clozapine users to derive reliable estimates of the real-world burden of clozapine adverse drug events (ADEs). A total of 2,659 participants in the FinnGen biobank project had a schizophrenia diagnosis and clozapine purchases with longitudinal electronic health record follow-up for up to 25 years after clozapine initiation. Diseases and health-related events enriched during clozapine use were identified, adjusting for disease severity. The incidence and recurrence of ADEs over years of clozapine use, their effect on clozapine discontinuation and deaths, and their pharmacogenetics were studied. Median follow-up time after clozapine initiation was 12.7 years. Across 2,157 diseases and health-related events, 27 were enriched during clozapine use, falling into five disease categories: gastrointestinal hypomotility, seizures, pneumonia, other acute respiratory tract infections, and tachycardia, along with a heterogeneous group including neutropenia and type 2 diabetes, among others. Cumulative incidence estimates for ileus (severe gastrointestinal hypomotility) and pneumonia were 5.3% and 29.5%, respectively, 20 years after clozapine initiation. Both events were significantly associated with increased mortality among clozapine users (ileus: odds ratio=4.5; pneumonia: odds ratio=2.8). Decreased genotype-predicted CYP2C19 and CYP1A2 activities were associated with higher pneumonia risk. Clozapine-induced ileus and pneumonia were notably more frequent than has previously been reported and were associated with increased mortality. Two CYP genes influenced pneumonia risk. Pneumonia and ileus call for improved utilization of available preventive measures.

Abstract P134: Effects of Antihypertensive Medications on All-cause Dementia: Evidence from Target Trial Emulation of 3 Million US Veterans

Introduction: Hypertension is a modifiable risk factor for all-cause dementia. Mechanistic and observational evidence suggest angiotensin receptor blockers (ARBs) and angiotensin converting enzyme inhibitors (ACEIs) may differentially affect dementia risk. Objective and Methods: We used longitudinal electronic health records of over 3 million US Veterans to conduct a target trial emulation with a new-user, active comparator design. We compared the intention-to-treat effect of initiating ARBs vs ACEIs (1/1/2000-12/31/2022) on dementia risk. We controlled for &gt;60 baseline covariates using inverse probability (IP) of treatment and overlap weighting. Dementia was ascertained using diagnosis codes and medications. We nonparametrically estimated the cause-specific cumulative incidence for each treatment group under the competing risk of death then computed treatment effects as weighted risk ratios with bootstrapped 95% confidence intervals. We also evaluated all-cause death and a composite outcome of dementia or death (dementia/Death) using Cox regression. Results: Among the 3,190,356 included Veterans (mean age 63, 5% female, 64% White, and 15% Black), 13% initiated ARBs and 87% initiated ACEIs. Overall, 8% developed dementia and 35% died over a median follow-up of 8 years. The risk of dementia was lower among ARB vs ACEI initiators (10-year risk ratio: 0.95 (0.94, 0.96)). The risk of all-cause death and dementia/Death were also lower among ARB vs ACEI initiators with overlap weighting. Conclusions: Among US Veterans with hypertension, ARB vs ACEI initiation was associated with a moderately lower risk of dementia. Future work assessing sustained treatment with ARB vs ACEI is needed to inform their use to reduce dementia risk.

Navigating a "Good Death" During COVID-19: Understanding Real-Time End-of-Life Care Structures, Processes, and Outcomes Through Clinical Notes.

The coronavirus disease 2019 (COVID-19) pandemic severely disrupted hospice care, yet there is little research regarding how widespread disruptions affected clinician and family decision-making. We aimed to understand how the pandemic affected structures, processes, and outcomes of end-of-life care. Retrospective narrative chart review of electronic health records of 61 patients referred and admitted to hospice from 3 New York City geriatrics practices who died between March 1, 2020, and March 31, 2021. We linked longitudinal, unstructured medical, and hospice electronic health record notes to create a real-time, multiperspective trajectory of patients' interactions with providers using directed content analysis. Most patients had dementia and were enrolled in hospice for 11 days. Care processes were shaped by structural factors (staffing, supplies, and governmental/institutional policies), and outcomes were prioritized by care teams and families (protecting safety, maintaining high-touch care, honoring patient values, and supporting patients emotionally and spiritually). Processes used to achieve these outcomes were decision-making, care delivery, supporting a "good death," and emotional and spiritual support. Care processes were negotiated throughout the end of life, with clinicians and families making in-the-moment decisions. Some adaptations were effective but also placed extraordinary pressure on paid and family caregivers. Healthcare teams' and families' goals to meet patients' end-of-life priorities can be supported by ongoing assessment of patient goals and process changes needed to support them, stronger structural supports for paid and family caregivers, incentivizing relationships across primary care and hospice teams, and extending social work and spiritual care.

Deep learning-based prediction of Clostridioides difficile infection caused by antibiotics using longitudinal electronic health records

Clostridioides difficile infection (CDI) is a major cause of antibiotic-associated diarrhea and colitis. It is recognized as one of the most significant hospital-acquired infections. Although CDI can develop severe complications and spores of Clostridioides difficile can be transmitted by the fecal-oral route, CDI is occasionally overlooked in clinical settings. Thus, it is necessary to monitor high CDI risk groups, particularly those undergoing antibiotic treatment, to prevent complications and spread. We developed and validated a deep learning-based model to predict the occurrence of CDI within 28 days after starting antibiotic treatment using longitudinal electronic health records. For each patient, timelines of vital signs and laboratory tests with a 35-day monitoring period and a patient information vector consisting of age, sex, comorbidities, and medications were constructed. Our model achieved the prediction performance with an area under the receiver operating characteristic curve of 0.952 (95% CI: 0.932–0.973) in internal validation and 0.972 (95% CI: 0.968–0.975) in external validation. Platelet count and body temperature emerged as the most important features. The risk score, the output value of the model, exhibited a consistent increase in the CDI group, while the risk score in the non-CDI group either maintained its initial value or decreased. Using our CDI prediction model, high-risk patients requiring symptom monitoring can be identified. This could help reduce the underdiagnosis of CDI, thereby decreasing transmission and preventing complications.

On the evaluation of synthetic longitudinal electronic health records

BackgroundSynthetic Electronic Health Records (EHRs) are becoming increasingly popular as a privacy enhancing technology. However, for longitudinal EHRs specifically, little research has been done into how to properly evaluate synthetically generated samples. In this article, we provide a discussion on existing methods and recommendations when evaluating the quality of synthetic longitudinal EHRs.MethodsWe recommend to assess synthetic EHR quality through similarity to real EHRs in low-dimensional projections, accuracy of a classifier discriminating synthetic from real samples, performance of synthetic versus real trained algorithms in clinical tasks, and privacy risk through risk of attribute inference. For each metric we discuss strengths and weaknesses, next to showing how it can be applied on a longitudinal dataset.ResultsTo support the discussion on evaluation metrics, we apply discussed metrics on a dataset of synthetic EHRs generated from the Medical Information Mart for Intensive Care-IV (MIMIC-IV) repository.ConclusionsThe discussion on evaluation metrics provide guidance for researchers on how to use and interpret different metrics when evaluating the quality of synthetic longitudinal EHRs.

Polygenic risk scores as a marker for epilepsy risk across lifetime and after unspecified seizure events

A diagnosis of epilepsy has significant consequences for an individual but is often challenging in clinical practice. Novel biomarkers are thus greatly needed. Here, we investigated how common genetic factors (epilepsy polygenic risk scores, [PRSs]) influence epilepsy risk in detailed longitudinal electronic health records (EHRs) of > 700k Finns and Estonians. We found that a high genetic generalized epilepsy PRS (PRSGGE) increased risk for genetic generalized epilepsy (GGE) (hazard ratio [HR] 1.73 per PRSGGE standard deviation [SD]) across lifetime and within 10 years after an unspecified seizure event. The effect of PRSGGE was significantly larger on idiopathic generalized epilepsies, in females and for earlier epilepsy onset. Analogously, we found significant but more modest focal epilepsy PRS burden associated with non-acquired focal epilepsy (NAFE). Here, we outline the potential of epilepsy specific PRSs to serve as biomarkers after a first seizure event.